Out-of-Distribution Generalization in Graph Foundation Models

Graphs are a fundamental data structure for representing relational information in domains such as social networks, molecular systems, and knowledge graphs. However, graph learning models often suffer from limited generalization when applied beyond their training distributions. In practice, distribution shifts may arise from changes in graph structure, domain semantics, available modalities, or task formulations. To address these challenges, graph foundation models (GFMs) have recently emerged, aiming to learn general-purpose representations through large-scale pretraining across diverse graphs and tasks. In this survey, we review recent progress on GFMs from the perspective of out-of-distribution (OOD) generalization. We first discuss the main challenges posed by distribution shifts in graph learning and outline a unified problem setting. We then organize existing approaches based on whether they are designed to operate under a fixed task specification or to support generalization across heterogeneous task formulations, and summarize the corresponding OOD handling strategies and pretraining objectives. Finally, we review common evaluation protocols and discuss open directions for future research. To the best of our knowledge, this paper is the first survey for OOD generalization in GFMs.

💡 Research Summary

Graphs are a universal data structure for representing relational information in domains ranging from social networks to molecular systems and knowledge graphs. While graph neural networks (GNNs) have become the workhorse for tasks such as node classification, link prediction, and graph‑level prediction, they typically overfit to the specific distribution of the training graph and struggle when deployed on new graphs whose topology, feature distribution, semantics, or task formulation differ. This paper surveys recent progress on Graph Foundation Models (GFMs) from the perspective of out‑of‑distribution (OOD) generalization, a viewpoint that has been largely missing from prior surveys that focus on architecture or pretraining objectives alone.

The authors first identify four axes of distribution shift that can affect graph learning: (1) structural shifts (changes in topology, motif frequencies, or node attributes), (2) domain shifts (dataset‑specific collection or annotation biases), (3) modality shifts (presence, quality, or alignment of auxiliary modalities such as text or molecular descriptors), and (4) task shifts (different supervision granularity or output spaces). They formalize OOD generalization by introducing latent generative factors Φ = (Φ_struct, Φ_dom, Φ_mod, Φ_task) that govern the data‑generating process in each environment. Training data are drawn from a mixture distribution p_src, while the test environment follows p_tgt; the goal is to learn a predictor f_θ that minimizes expected loss under p_tgt despite the shift in Φ.

Based on whether the downstream task specification Φ_task is assumed fixed or allowed to vary, the survey organizes existing GFMs into two broad categories.

1. Homogeneous‑Task GFMs (fixed task formulation).
These models target a single downstream task (e.g., node classification) throughout pretraining and inference. OOD robustness is achieved by learning representations that are invariant to structural, domain, or modality changes while keeping the supervision unchanged. Representative approaches include:

• GraphFM – tokenizes arbitrary graphs into a fixed latent token set, pretrains on >100 datasets with a unified node‑classification loss, and adapts only lightweight dataset‑specific heads.
• AnyGraph – employs a mixture‑of‑experts backbone with dynamic routing based on self‑supervised link‑prediction signals, enabling domain‑specific expert specialization without global parameter updates.
• MDGPT – introduces domain tokens and alignment functions that modulate node features before feeding them to a shared encoder; a universal link‑prediction objective enforces a consistent task interface across domains.
• PatchNet – learns learnable graph patches that reorganize heterogeneous node attributes into a common patch space, mitigating feature mismatches across domains.

Other methods such as GraphAny, GraphLoRA, MDGFM, SAMGPT, GraphCLIP, RiemannGFM, and GOODFormer rely on multi‑graph pretraining, contrastive alignment, invariant subgraph modeling, or Bayesian in‑context inference to enforce structural or domain invariance. The common theme is to expose the model to diverse graphs during pretraining, use loss functions that encourage stable latent factors, and keep downstream adaptation lightweight.

2. Heterogeneous‑Task GFMs (task‑varying).
These models explicitly handle variability in the downstream task definition, often through prompting, unified tokenization, or multimodal alignment. Key examples are:

• OFA – adopts a unified semantic prompting scheme that encodes task instructions together with graph inputs, allowing the same model to perform classification, regression, or QA by simply changing the prompt.
• LLaGA – aligns structural and semantic information across graph and text modalities, training with a likelihood maximization objective that supports question‑answering style tasks.
• OpenGraph, GOFA, LLM‑BP, GIT, GFT, AutoGFM, UniGraph, and UniGraph2 – employ next‑token prediction, tree‑structured reconstruction, architecture‑aware contrastive learning, or masked modeling on a unified graph‑text token stream, thereby learning task‑agnostic representations that can be steered by task‑specific prompts at inference time.

These approaches shift the burden of task adaptation from parameter fine‑tuning to prompt engineering, enabling rapid deployment on unseen tasks without catastrophic forgetting.

The survey also reviews evaluation protocols used to assess OOD performance. Benchmarks typically vary one factor at a time: structural OOD is tested by altering graph size, density, or motif distribution; domain OOD by cross‑dataset validation; modality OOD by dropping or corrupting auxiliary modalities; and task OOD by swapping downstream objectives. The authors note that most existing works focus on structural and domain shifts, while systematic studies of modality and task shifts remain scarce.

Finally, the paper outlines open research directions: (i) developing quantitative metrics and visualizations for the latent factors Φ; (ii) designing encoders that jointly model multi‑domain and multi‑modality interactions; (iii) providing theoretical foundations for prompt‑based task transfer and efficient prompt generation; and (iv) integrating real‑time OOD detection with adaptive model updates.

In summary, this survey provides the first comprehensive treatment of OOD generalization in graph foundation models, clarifying how different design choices address specific distribution shifts, summarizing current strengths and gaps, and charting a roadmap for future advances in robust, versatile graph AI.